New: independent coding checks for insurers and providers. See how Lens works →

Back

RADV Audits: Sampling, Extrapolation, and Chart Defense

· 8 min read
Risk adjustment

A $60,000 RADV sample error can become a $30 million extrapolation. See how CMS's 90% confidence rule works and how to build a chart defense.

A RADV audit is how CMS checks whether the diagnoses a Medicare Advantage (MA) organization submitted for payment are supported by medical records. Risk Adjustment Data Validation runs at the contract level: CMS samples enrollees, reviews the records behind their hierarchical condition categories (HCCs), recalculates their risk scores, and recovers the difference.

The sample is small. The exposure isn't. Under the method CMS has published, an error found in a few dozen charts can be extrapolated across every enrollee in the contract. That's the part finance leaders lose sleep over, and it's the part this article explains in detail, along with where the legal fight over it stands and how to defend the charts that drive it.

The timing is live. For payment year 2024, CMS's medical record submission window opened September 18, 2026 and closes February 5, 2027.

Where RADV stands at the end of 2026

Date Event
January 2023 CMS finalizes a RADV rule allowing extrapolation from payment year 2018 and dropping the fee-for-service (FFS) adjuster
May 21, 2025 CMS announces it will audit all eligible contracts every year, grow its coder workforce from 40 to about 2,000, and finish payment years 2018 to 2024 by early 2026
September 25, 2025 A federal court in Texas vacates the 2023 rule on procedural grounds (Humana v. Becerra)
November 2025 HHS appeals to the Fifth Circuit
January 2026 CMS memo: it will comply with the order while it stands, and continue audits
September 2026 PY2024 audits open; enrollee lists published September 11
October 5, 2026 Fifth Circuit hears argument in Humana v. Kennedy; no decision as of November 5, 2026

Two points get lost in the headlines. The district court didn't hold that extrapolation is unlawful; it held that CMS didn't follow proper notice-and-comment procedure. And CMS hasn't stopped auditing. Its PY2024 audit methods guidance acknowledges the vacated rule and the appeal, collects only the sampled enrollees' errors for now, and reserves the right to collect extrapolated amounts later where legally permitted.

Plan as if extrapolation comes back. If it does, the audits already under way are the ones it applies to.

How CMS samples a contract

The unit of the sample is the enrollee, not the HCC. CMS ranks contracts by the size of their sampling frame, splits them into four strata, and draws a simple random sample without replacement from each contract.

Stratum Which contracts Enrollees sampled
1 The 10 largest contracts 200
2 Largest third of the rest 100
3 Middle third 50
4 Smallest third 35

If a contract's sampling frame is no bigger than its stratum's sample size, every enrollee in the frame is audited. For each sampled enrollee, CMS audits the HCCs that were used for payment, using records with dates of service in the data collection year (calendar year 2023 for PY2024).

What you can submit

  • At least one valid medical record per audited HCC, and no more than twice the number of audited HCCs per enrollee. An enrollee with three audited HCCs gets up to six records.
  • CMS-generated attestations can cure a missing signature or credential, but only for physician and hospital outpatient records, not inpatient records. They must be signed by the practitioner who saw the patient, and an altered or mismatched attestation invalidates the record.
  • Only valid records move on to diagnosis abstraction, and only valid records can support an appeal.

The two-records-per-HCC limit is the most important tactical fact in the guidance. Every HCC gets at most two chances. Picking the right encounter matters as much as having one.

How extrapolation works

For each sampled enrollee, CMS recalculates the risk score using only the diagnoses it can abstract from valid records, converts that to monthly payments using the county rates, and subtracts it from what was paid. The result is the enrollee's payment error. It can be positive, zero or negative.

CMS then uses the average payment error across the sample and its variance to build a 90 percent confidence interval. Under the PY2024 method, CMS extrapolates only when the lower bound of that interval is above zero. If the lower bound is zero or below, it doesn't. Frames under 30 enrollees are audited in full without extrapolation, and a net underpayment doesn't generate any payment back to the plan.

A worked example

Illustrative figures, not from a real contract.

A Stratum 2 contract has 50,000 enrollees in its sampling frame. CMS samples 100. Across those 100, the average payment error is $600 per enrollee, with a standard deviation of $2,400.

Step Calculation Result
Sample-only recovery 100 × $600 $60,000
Standard error $2,400 ÷ √100 $240
Lower bound of 90% interval $600 − (1.660 × $240) $201.50
Lower bound above zero? Yes Extrapolation allowed
Extrapolated to the frame $600 × 50,000 $30,000,000

A $60,000 sample finding becomes a $30 million exposure. That 500-to-1 ratio is just the frame size divided by the sample size, which is why large contracts carry the most risk even with the largest samples.

Now change one number. If the same $600 average came with a standard deviation of $5,000, because errors were concentrated in a few high-value enrollees, the lower bound would fall to about −$230. No extrapolation. Variance protects a contract only when errors are lumpy and rare. A program with small, consistent errors across many enrollees is the profile that extrapolates.

Why the FFS adjuster mattered

The CMS-HCC model is calibrated on Original Medicare data, which contains its own documentation errors. For years, CMS's RADV methodology included an FFS adjuster to offset that baseline, so plans weren't penalized for error rates the model already assumed. The 2023 rule removed it, which raised extrapolated amounts. That removal, along with how CMS finalized it, is at the center of the Humana litigation. The PY2024 guidance doesn't apply an FFS adjuster.

Chart defense: winning each HCC

RADV is won or lost one HCC at a time. Every HCC either validates or it doesn't, and each failure feeds the average error that drives extrapolation.

Common failure Why it fails Prevent it before the audit Fix it at audit time
Diagnosis only in the problem list or history Not shown as addressed at an encounter Documentation standards and coder rules Pick an encounter where the condition was assessed
Missing signature or credentials Record not authenticated EHR signature controls CMS-generated attestation (physician and outpatient only)
Wrong provider type or specialty Not an acceptable risk adjustment source Source rules at submission Find a qualifying encounter instead
Date outside the collection year Doesn't support the payment year Date checks at submission Choose an in-year encounter
Severity or type not stated by the clinician Coded beyond the documentation Query during the concurrent window Choose the record with the specific statement
Acute condition coded in an office setting History of the condition coded as acute Edits on acute-only diagnoses outside hospital records None; the HCC usually fails

What the record has to show is covered in depth in MEAT documentation that matches RADV. The short version: a clear, authenticated, in-year, face-to-face record from an acceptable provider, where the clinician addresses the condition at the specificity coded.

Choosing the best record

With at most two records per HCC, record selection is a coding skill, not a clerical task. A credentialed coder should pick records using a fixed order of preference:

  1. An encounter where the condition is assessed and managed, not just listed
  2. With a valid signature and credentials already on the note
  3. From the most clinically relevant provider (the treating specialist for a specialist-managed condition)
  4. Stating the type, stage or severity that matches the HCC submitted

When no record clears the bar, don't submit a weak one hoping it passes. An invalid record can't support an appeal, and it signals a pattern to reviewers.

If you have a PY2024 audit open now

With the submission window closing February 5, 2027, the order of work matters more than the volume.

  1. Triage the enrollee list by HCC value. Start with enrollees whose audited HCCs carry the most weight. Each one moves the average payment error, and the average is what extrapolates.
  2. Pull every in-year encounter, then pick two. Retrieve all 2023 encounters for each audited HCC before choosing, so the two you submit are the strongest, not the first found.
  3. Fix authentication early. Identify records missing signatures or credentials in the first weeks and request CMS-generated attestations from the treating practitioners while there's time. Inpatient records can't be cured this way.
  4. Second-review every package. A credentialed coder who didn't select the records checks each one against the chart defense table above before it's submitted.
  5. Log what fails. Every HCC you can't support is a pattern to fix in current-year coding, and a candidate for the delete process.
  6. Ask early if you'll need a hardship exception. The request deadline is February 19, 2027, and it shouldn't be the first time the team discusses one.

Know your number before CMS does

The best RADV defense is done a year earlier. Run CMS's own method on your own contract:

  1. Draw a random sample of enrollees from your frame, sized to your stratum.
  2. Have credentialed coders who didn't code the original charts validate every payment HCC against the records you'd actually submit.
  3. Recalculate each enrollee's score with only the validated HCCs. The arithmetic is the same one shown in how each coefficient builds a RAF score.
  4. Compute the average payment error, its standard deviation and the 90 percent lower bound.

That gives you an estimate of your extrapolation exposure, and a list of the HCC patterns causing it.

Then act on what you find. Unsupported HCCs should be deleted, and identified overpayments reported and returned under the MA overpayment rules. A delete process is also the single clearest signal of good faith. Programs that only ever add HCCs are the profile that RADV, OIG and the Department of Justice look for first.

The other audit track: OIG

CMS isn't the only auditor. The HHS Office of Inspector General runs its own audits of MA organizations, usually targeting high-risk diagnosis groups, and recommends recoveries.

The pattern is consistent: acute conditions coded from physician office records. In a May 2026 report, OIG found that acute stroke diagnoses on physician records were unsupported for all 97 enrollees it sampled, and estimated $462 million in potential overpayments for 2021. It recommended that CMS block acute stroke diagnoses from physician records when there's no matching hospital record in the same year.

If your program submits acute diagnoses such as stroke from office visits, that's the first place to audit.

Appeals

MA organizations can appeal RADV findings under 42 CFR 422.311, covering both medical record review determinations and payment error calculations. Two limits matter. Only valid medical records can support a record review appeal, so a record rejected as invalid can't be rescued on appeal. And CMS's decisions on whether to credit additional HCCs found in the records can't be appealed.

FAQ

What is a RADV audit?

A Risk Adjustment Data Validation audit is CMS's review of the diagnoses a Medicare Advantage organization submitted for payment. CMS samples enrollees in a contract, checks each payment HCC against medical records, recalculates risk scores, and recovers overpayments.

How many enrollees does a RADV audit sample?

Under CMS's PY2024 method, 35 to 200 enrollees per contract, depending on contract size: 200 for the 10 largest contracts, then 100, 50 or 35 for the rest by size. Small contracts may be audited in full.

Can CMS extrapolate RADV findings?

CMS's published method allows extrapolation when the lower bound of a 90 percent confidence interval around the average payment error is above zero. The 2023 rule behind it was vacated in September 2025 and is on appeal. For PY2024, CMS is collecting sampled enrollees' errors and reserving extrapolation where legally permitted.

How many medical records can be submitted per HCC?

At least one valid record per audited HCC, and up to twice the number of audited HCCs per enrollee. CMS-generated attestations can fix a missing signature on physician and outpatient records, but not inpatient records.

When are PY2024 RADV records due?

CMS's PY2024 guidance sets the medical record submission deadline at February 5, 2027, 11:59 p.m. Eastern, with hardship exception requests due February 19, 2027.

This article is general education on RADV audits. It isn't legal advice on audit exposure or a coding determination for any specific patient or contract.

More from the Journal

All articles

See it on your own charts.